Abstract
Models trained on protein sequences can predict mutational effects, but their outputs represent a convolution of different phenotypes. Isolating stability-independent activity effects is essential for rational enzyme design, allosteric drug design and functional annotation, yet existing pLM-based deconvolution methods lack interpretability. Here, we introduce a hybrid framework that pairs stability predictors with probabilistic graphical models of coevolution. Benchmarked on all enzymes with available multidimensional deep mutational scanning data, our framework isolates stability-independent activity effects with a fidelity that matches analogous pLM-based approaches. Crucially, we reveal biophysically plausible communication pathways that align with previous knowledge and that we validate with experimental data across a diverse set of targets. This hybrid approach merges statistical and geometric deep learning into a zero-shot, explainable tool for targeted engineering and mechanistic discovery.